Tree of Concepts: Interpretable Continual Learners in Non-Stationary Clinical Domains

The harder problem Tree of Concepts is solving isn't forgetting, which is a well-studied failure mode, but rather explanation drift: when a model updates on new patient data, the concepts it uses to justify a decision can shift even if accuracy holds, which is a regulatory and liability problem that pure performance metrics won't catch.

This sits in a cluster of interpretability work Modelwire has been tracking across the past week. The ORCA paper from April 16 approached interpretability as a post-training audit problem for SVMs, keeping the model fixed and making its structure legible after the fact. Tree of Concepts takes the opposite stance: interpretability is a constraint baked into the training loop itself, not a lens applied afterward. That distinction matters clinically because post-hoc explanations can become stale or inconsistent as a model continues learning, whereas a concept bottleneck architecture forces the system to route predictions through human-readable intermediate representations at every update. Neither paper addresses the other's failure mode, which is worth holding in mind.

The real test is whether the decision tree interface remains stable across genuinely adversarial distribution shifts, such as a formulary change or a new patient population. If the authors release a clinical benchmark with documented shift events and the concept labels stay consistent across those boundaries, the interpretability claim is substantive; if evaluation stays on controlled synthetic splits, it isn't.